Abstract
We describe ongoing research in developing audio classification systems that use a spiking silicon cochlea as the front end. Event-driven features extracted from the spikes are fed to deep networks for the intended task. We describe a classification task on naturalistic audio sounds using a low-power silicon cochlea that outputs asynchronous events through a send-on-delta encoding of its sharply-tuned cochlea channels. Because of the event-driven nature of the processing, silences in these naturalistic sounds lead to corresponding absence of cochlea spikes and savings in computes. Results show 48% savings in computes with a small loss in accuracy using cochlea events.